STAR: croSs-plaTform App Recommendation Da Cao, Xiangnan He, - - PowerPoint PPT Presentation

STAR: croSs-plaTform App Recommendation

Da Cao, Xiangnan He, Liqiang Nie, Xiaochi Wei, Xia Hu, Shunxiang Wu, Tat-Seng Chua

This work was finished when Da Cao was a Ph.D. student in XMU and a visiting student in NUS. Currently Da Cao is an assistant professor in HNU. caoda0721@gmail.com

7/16/2018 1

Submit: June 2016 Accepted: Nov 2016 Published: July 2017

Outline

• Background
• Proposed Method
• Experiments and Results
• Conclusion

7/16/2018 2

App Development

Mobile Network

7/16/2018 3

App-Driven Life Multi-Platform Overwhelmed

Two App Recommendation solutions

Single Platform Cross-Platform

Who wins?

4 7/16/2018

VS Russia Croatia

Challenges

• Platform Variance
• Data Heterogeneity
• Data Sparsity
• Cold-Start Problem

7/16/2018 5

Outline

• Background
• Proposed Method
• Experiments and Results
• Conclusion

7/16/2018 6

Proposed Method

7/16/2018 7

Cold-Start Problems

New-user cold-start New-App cold-start

7/16/2018 8

The question mark “?” stands for the ratings that we wish to predict, and the label “new” means the user or App is new to the platform and has no rating history on it.

Outline

• Background
• Proposed Method
• Experiments and Results
• Conclusion

7/16/2018 9

Dataset and Evaluation

7/16/2018 10

We selected users who rated at least once on both of these platforms.

We selected users who had at least two ratings on all platforms.

Rating Prediction

• MAE
• RMSE

Top-N Recommendation

• Recall
• NDCG

Research Questions

• (RQ1). How does STAR perform as compared to other state-of-the-art competitors?
• (RQ2). How is the performance of STAR in handling the new-user and new-App cold-start

problems?

• (RQ3). Whether the rated App on current platform is the user's preferable one as compared

to the same App on other unrated platforms?

• (RQ4). How do the common features and specific features of Apps contribute to the overall

effectiveness of STAR?

• (RQ5). In addition to rating prediction that is prevalent to a recommendation algorithm, how

does STAR perform in the more practical top-N recommendation?

7/16/2018 11

Baseline Methods

• SVD++ [Koren 2008] (Collaborative Filtering)
• RMR [Ling et al. 2014] (Semantics Enhanced Recommendation)
• CTR [Wang and Blei 2011] (Semantics Enhanced Recommendation)
• FM [Rendle et al. 2011] (Context-Aware Recommender System)
• CMF [Singh and Gordon 2008] (Cross-Domain Recommender System)
• WMF [Hu et al. 2008] (Collaborative Filtering)
• Popular (Non-personalized method)

7/16/2018 12

Overall Performance Comparisons (RQ1)

7/16/2018 13

Handling Cold-Start Problems (RQ2)

7/16/2018 14

User Preference on App-Platform (RQ3)

7/16/2018 15

• 1. Our method favors the current platform better.
• 2. The gap between rating predictions of current platform and other platforms on

the iPhone-iPad dataset is larger than that of the iPhone-iPad-iMac dataset.

Justification of Common Features and Specific Features (RQ4)

7/16/2018 16

Evaluation of Top-N Recommendation (RQ5)

7/16/2018 17

Outline

• Background
• Proposed Method
• Experiments and Results
• Conclusion

7/16/2018 18

Challenges Solved

• Platform Variance
• Data Heterogeneity
• Data Sparsity
• Cold-Start Problem

7/16/2018 19

Website

Data & code are available at http://apprec.wixsite.com/star

7/16/2018 20

Thanksgiving

7/16/2018 21

Xiangnan He (NUS) Liqiang Nie (SDU) Xiaochi Wei (BIT) Xia Hu (Texax A&M) Shunxiang Wu (XMU) Tat-Seng Chua (NUS)

Co-authors: And all audiences…